IMPLEMENTATION OF FUZZY C-MEANS CLUSTERING AND PROXIMITY-IMPACT-POPULARITY ON USER-BASED COLLABORATIVE FILTERING

RUBILA DWI ADAWIYAH , Arif Nurwidyantoro, S.Kom, M.Cs

2018 | Skripsi | S1 ILMU KOMPUTER

Memory-based Collaborative Filtering (CF) adalah metode sistem rekomendasi yang banyak digunakan karena implementasinya yang mudah. Namun, memory-based CF memiliki permasalahan sparsity, skalabilitas, dan cold-start. Masalah-masalah ini mempengaruhi performa sistem rekomendasi. Penelitian ini mencoba untuk mengatasi skalabilitas dan masalah sparsity pada salah satu tipe memory-based CF yaitu user-based CF dengan mengimplementasikan Fuzzy C-Means (FCM) dan mengatasi masalah cold-start dengan menggunakan Proximity-Impact-Popularity (PIP) sebagai perhitungan similaritas. Ada dua metode defuzzifikasi yang digunakan yaitu Best Cluster dan All Cluster defuzzification. Dataset yang digunakan yaitu MovieLens dataset dengan tingkat sparsity 93,69%. Hasil dari penelitian ini adalah PIP menghasilkan akurasi dan coverage yang lebih tinggi (MAE = 0,7734 dan coverage = 74,32%) daripada korelasi Pearson (MAE = 0,849588 dan coverage = 47%). Namun, skalabilitas PIP sangat buruk (throughput = 4239,113 rec/sec). Implementasi FCM pada kedua metode defuzzifikasi tersebut meningkatkan skalabilitas sistem dengan throughput yang lebih tinggi (throughput FCM Best Cluster defuzzification=7386,41 rec/sec, throughput FCM All Cluster defuzzification=15552.1 rec/sec). Dalam peningkatan jumlah cold-user (pengguna baru), FCM dengan kedua metode defuzzifikasi juga menghasilkan akurasi dan coverage yang lebih tinggi.

Memory-based Collaborative Filtering (CF) is a widely used recommender system method due to its easy implementation. However, it suffers from sparsity, scalability, and cold start problems. These problems influence the performance of the recommender system. This research attempted to overcome scalability and sparsity problem on one type of memory-based CF which is user-based CF by implementing Fuzzy C-Means (FCM) clustering and to overcome cold-start problem by using Proximity-Impact-Popularity (PIP) as similarity measure. There are two defuzzification methods namely Best Cluster and All Cluster defuzzification. The system is implemented in MovieLens dataset with sparsity level 93,69 %. The result is PIP gave higher accuracy and higher coverage (MAE = 0,7734 and coverage = 74.32%) than Pearson correlation (MAE = 0.849588 and coverage = 47%). However, it still suffers very poor scalability (throughput = 4239,113 rec/sec). The implementation of FCM both defuzzification methods improve the scalability of the system with higher throughput (throughput FCM Best Cluster defuzzification = 7386,41 rec/sec, throughput FCM All Cluster defuzzification = 15552,1 rec/sec). In increasing number of cold users, FCM with both defuzzification methods also result in higher accuracy and coverage.

Kata Kunci : Collaborative Filtering, Recommender System, Sparsity, Cold Users, Fuzzy C-Means